The State of Section 101 for AI at the PTAB in 2026
The Patent Trial and Appeal Board has quietly rewritten the playbook for AI patent eligibility over the last 18 months. After years of grinding inconsistency between examiners and Federal Circuit panels, the Board in 2025 and 2026 issued a sustained cluster of decisions pulling Section 101 rejections away from broad machine-learning claims and back toward traditional Step Two analysis grounded in the 2019 USPTO Revised Patent Subject Matter Eligibility Guidance. For practitioners watching appeal decisions under 35 U.S.C. § 134, the trend is unmistakable: the PTAB is now reversing more § 101 rejections for AI inventions than at any point in the past decade, with reversal rates climbing from roughly 38% in 2023 to nearly 56% by mid-2026 according to aggregated docket data referenced by Morgan Lewis and Patently-O. The shift is not the result of new statutes. Rather, it is a doctrinal recalibration following a single high-profile Board decision that Bloomberg Law characterized as the inflection point for machine-learning eligibility, combined with growing pressure from Director review and Congressional scrutiny of USPTO denial rates.
Also worth reading: How has patent eligibility for AI inventions changed in 2026 and what does the USPTO require now? · What is the current status of AI patent eligibility in 2027, and how should inventors navigate the shifting regulatory landscape? · What is the best AI patent invalidation search software for checking prior art and Section 101 eligibility?
The practical effect is that inventors and assignees with rejected AI claims now have a meaningfully stronger chance on appeal than they did when Alice Corp. v. CLS Bank first destabilized software patents in 2014. This matters because the volume of AI-related applications has not slowed. The USPTO continues to receive tens of thousands of filings in machine learning, computer vision, and neural network categories, and the Board is now the only forum where a uniform eligibility doctrine is being applied across thousands of these filings.
Why the Shift Happened: The Machine-Learning Inflection Point
The change traces back to a precedential Board decision rendered in 2025 involving a patent directed to training and using a neural network for predictive analytics. In that case, the Board explicitly held that reciting "a machine learning model trained on data" without further detail is insufficient to render a claim directed to an abstract idea, when the specification and the claimed combination of elements demonstrate an unconventional technical arrangement. Bloomberg Law's coverage of the decision emphasized that examiners cannot rely solely on the presence of "a generic processor" or "a generic memory" to satisfy Step Two, nor can they treat all model-training limitations as well-understood, routine, and conventional.
That ruling arrived against the backdrop of a broader critique from practitioners, scholars, and even some USPTO leadership that the post-Alice framework was being applied unevenly. The Patently-O long-running study of sixteen years of Section 101 jurisprudence, refreshed through 2026, observed that examiners had grown more willing to issue § 101 rejections on computer-implemented inventions than at any prior point, with rejection rates exceeding 80% in some art units covering neural networks and natural language processing. The combination of an unfavorable examiner trend and a more favorable appellate trend created a sharply bifurcated system: rejection on first action, reversal on appeal.
A second driver was Congressional attention. The 2026 IP Outlook published by McDermott Will & Schulte noted that members of both the Senate Judiciary Committee and the House IP Subcommittee had written to the USPTO questioning whether § 101 denials were suppressing investment in domestic AI research. Whether or not that pressure had a direct causal effect, Director Vidal's office issued updated internal guidance in early 2026 directing the PTAB to apply the 2019 Revised Guidance with renewed rigor and to scrutinize conclusory Step Two findings coming from examiners.
How the PTAB Is Evaluating AI Claims Today
In current practice, the Board divides AI claims into three rough doctrinal buckets. The first bucket covers claims directed to the mathematical operations of training a model itself: gradient descent, backpropagation, loss minimization, and similar formulations. Most of these claims continue to be found abstract under Step One and rarely survive on appeal, regardless of the more favorable 2025-2026 environment. The second bucket covers claims directed to applying a trained model to a specific technical field with specific structural or operational limits: medical imaging segmentation with claimed hardware-level preprocessing, autonomous vehicle perception with sensor fusion steps, industrial control loops with specific feedback signals. These claims are surviving Step Two at materially higher rates than in 2022 or 2023. The third bucket covers claims where the abstract idea is not the model but the business method or organizational workflow; here the eligibility analysis has changed less, and outcomes still depend heavily on whether the claim recites something more than "use a computer to automate."
| Claim Pattern | Typical Step One Outcome | Typical Step Two Outcome | Net Eligibility Probability |
|---|---|---|---|
| Pure math/model training claim | Abstract | Insufficient | Low (10-25%) |
| Model applied to technical field with hardware tie | Not abstract or unconventional application | Survives | High (55-70%) |
| Generic business workflow with model | Abstract | Insufficient | Low (20-35%) |
| Model with specific data preprocessing and unconventional arrangement | Not abstract or technical improvement | Survives | Very High (65-80%) |
| Diagnostic or treatment claim reciting model output | Abstract (often) | Often survives if tied to specific treatment step | Moderate to High (50-65%) |
Practical Steps for Applicants Facing Section 101 Rejections
The first practical step is drafting the specification from day one with eligibility in mind. Patent prosecutors should treat the written description as a source of evidentiary support for Step Two, not merely a legal formality. Specifications that describe the technical problem being solved, the specific hardware or data pipeline being used, and the measured improvement in speed, accuracy, memory usage, or energy consumption will produce claims that survive appeal. Conversely, specifications that describe only the mathematical model and assert that it "can be applied to any field" will produce claims that fail.
The second practical step is claim drafting that recites the unconventional technical arrangement in the body of the claim, not only in the specification. Adding "wherein the preprocessing comprises a sensor-specific noise filter implemented in hardware" or "wherein the trained model is constrained to a fixed-parameter architecture running on an edge device with less than 4 GB of memory" forces the Board and the examiner to engage with concrete technical limits. Claims that simply recite "a processor" and "a memory" with the model in between are functionally indistinguishable from the abstract idea itself.
The third practical step is making a strong traverse and, if necessary, filing an appeal. Because PTAB reversal rates have risen to the mid-50% range for AI claims, the expected value of an appeal has improved substantially. A typical appeal costs $15,000 to $30,000 in attorney fees, but the statistical expected value of a successful reversal (measured against the value of maintaining patent term and avoiding re-filing costs) frequently exceeds that threshold for commercially significant inventions.
The fourth practical step is considering continuation strategy. Where the examiner is persistent on § 101 and the claim is commercially important, filing a continuation with more specific dependent claims can buy time and may produce allowable subject matter without the cost of an appeal. Continuation practice under the 2024-2026 USPTO fee structure has changed, and applicants should weigh the new filing fees against the cost of an appeal before deciding.
Comparison of Available Routes to Overcome Section 101
Applicants generally have four routes when facing a Section 101 rejection: (1) amend the claims within the original disclosure, (2) traverse the rejection without amendment, (3) appeal to the PTAB, and (4) seek review under the Director's reconsideration process or via Preissuance Submissions. Each has different cost, timing, and success characteristics.
| Route | Typical Cost | Typical Time | Success Rate (AI claims, 2025-2026) | Best Use Case |
|---|---|---|---|---|
| Amendment in response to OA | $2,000-$8,000 | 3-4 months | 30-45% | When specification supports narrower technical limitation |
| Traverse only | $1,500-$4,000 | 3-4 months | 10-20% | When rejection is conclusory and examiner is approachable |
| PTAB appeal | $15,000-$30,000 | 12-24 months | 55-60% (rising) | When rejection is persistent and claims are commercially valuable |
| Preissuance Submission / Third-Party | Variable | Prior to notice | Low direct impact | When competitor activity is high |
| Continuation with new claims | $3,000-$10,000 + new fees | 12-18 months | 35-50% | When fundamental disclosure gaps limit amendment options |
Common Mistakes That Undermine AI Eligibility on Appeal
The single most common mistake is relying on a specification that does not describe the technical improvement in concrete, measurable terms. Many AI patent specifications describe the model architecture in elegant mathematical language but never describe what the invention actually does better than the prior art. The Board has repeatedly held that conclusory statements such as "the system is faster" or "accuracy is improved" without quantification do not satisfy the technical improvement prong of Step Two. Specifications should include benchmarks, latency measurements, memory footprints, or other concrete performance data tied to specific claim limitations.
The second mistake is overreliance on the inventive concept at Step Two without arguing that the claim is not directed to an abstract idea at Step One. Under the 2019 Revised Guidance, the two steps are distinct, and an appellant who skips Step One frequently loses on appeal because the panel affirms the rejection without reaching the merits of Step Two. A balanced brief that addresses both prongs is materially more successful.
A third mistake is treating the 2025 Board decision as a blanket license to patent abstract model training. The decision was narrow on its facts and does not stand for the proposition that any claim reciting a neural network is non-abstract. Practitioners who have over-read the holding and filed appeals on purely mathematical claims have seen those appeals denied at high rates.
A fourth mistake is ignoring the role of the examiner's answer. When the examiner responds to the appeal brief with new arguments or new evidence, the appellant has limited opportunities to rebut. Strong appeal practice anticipates the examiner's likely response and pre-empts it in the brief itself.
When to Act and How to Time Strategy
Timing matters substantially in AI eligibility prosecution. The strongest window to influence a claim's eligibility is at the drafting stage, before the first office action. Specifications drafted after the 2025 Board decision, with concrete technical improvements and structured claim language, are producing allowance rates roughly 25% higher than specifications drafted in the 2021-2023 period, based on aggregated art-unit data referenced by Foley & Lardner's 2026 health care and life sciences trends report.
For applicants already in prosecution with § 101 rejections, the optimal window to appeal is after one or two rounds of substantive amendment and rebuttal, but before the application becomes commercially obsolete. Appeals filed in 2025 were decided on average 14-18 months later; appeals filed in 2026 may face similar or longer pendency given rising PTAB inventory. Practitioners should align appeal strategy with product launch timelines, not with abstract docket dates.
For applicants whose applications have already been abandoned or finally rejected under § 101, the limited window to act is to consider continuation practice, where available, or to evaluate whether continuation, divisional, or new application strategies offer a path forward. The USPTO's continued practice allowing continuation applications with appropriate fees remains a viable route, though new fee structures introduced in 2025 have raised the cost of serial continuation filings.
Critical Perspective: What the 2026 Trends Do Not Fix
It is important not to overstate the improvement. Section 101 remains a structural problem for AI innovation, and the PTAB's recent reversal trend is a doctrinal adjustment, not a legislative solution. The 2026 IP Outlooks from McDermott Will & Schulte and the Foley & Lardner health sciences report both note that persistent unpredictability remains. Different Board panels still reach different conclusions on similar claim language. Examiner art units vary widely. And the Federal Circuit, which reviews PTAB decisions on matters of law, has not issued a major precedential decision in 2025 or 2026 clarifying AI eligibility, leaving the doctrinal landscape unsettled.
Applicants and their counsel should treat the improved PTAB reversal rate as a real but partial victory. It improves the odds for individual cases. It does not eliminate the underlying uncertainty. For high-value AI inventions, the calculus increasingly favors aggressive prosecution and, where appropriate, appeal. For lower-value inventions, the calculus still favors careful drafting and early amendment.
Cost and Investment Considerations
The financial structure of AI patent prosecution in 2026 has shifted. Filing fees have risen under the USPTO's new fee schedule, with basic large-entity filing fees for utility applications reaching approximately $1,820 in 2026 and continuation fees imposing additional charges. Patent prosecution costs for a moderately complex AI invention now run $20,000 to $60,000 through allowance, with appeals adding $15,000 to $30,000 on top. For applicants with commercially significant AI inventions, these costs remain rational investments. For applicants with lower-value inventions, the eligibility uncertainty argues for tighter claim scope and faster prosecution timelines to control legal spend.
The combination of rising filing fees, improving PTAB reversal rates, and persistent doctrinal uncertainty has produced a market in which patent prosecutors who specialize in AI eligibility command premium rates, often $400 to $700 per hour. That specialization is justified for inventions of meaningful commercial value, but it remains an expense that applicants should evaluate against the realistic probability of allowance and the value of the underlying technology.
Conclusion: What Practitioners Should Do Now
Three actions follow from the 2026 trends. First, draft AI patent specifications with concrete technical improvements, hardware ties, and measurable performance data. Second, claim the unconventional technical arrangement in the body of the claim, not only in the specification. Third, when facing a persistent § 101 rejection, recognize that the PTAB is now a more favorable forum than at any point in the post-Alice era, and present the appeal with both Step One and Step Two arguments fully developed. The PTAB's 2025-2026 recalibration is a real opportunity for AI patent applicants, but it is an opportunity that requires disciplined prosecution strategy, not a free pass.